Neural Image Enhancement Using Differential Images and Edge-Based Gain
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image processing technologies face challenges in achieving lightweight neural networks while improving image quality, and there is a difficulty in considering the features of images input to subsequent algorithms due to increased hardware complexity and lack of integration with previous algorithm outputs.
Innovation Solution
An image processing device and method that utilizes a neural network to generate a first differential image with higher resolution and tone curve parameters, applies gain values to filtered images based on edge maps, and combines these to enhance sharpness and contrast through a series of sub-neural networks and filtering processes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If traditional linear systems are used for image processing, then hardware complexity is low, but image quality improvement capability is limited
Solution Approach 1:
The patent segments the image processing task into multiple specialized neural network modules: a first neural network for noise cancellation, a second neural network for super-resolution, a third neural network for sharpness enhancement, a fourth neural network for contrast enhancement, and a fifth neural network for color expansion. Each module processes specific aspects of image quality independently, allowing the system to achieve comprehensive image enhancement while managing hardware complexity through modular design.
Solution Approach 2:
The patent creates a unified image processing system where multiple neural networks share common infrastructure and processing pipelines. The system integrates noise cancellation, super-resolution, sharpness enhancement, contrast enhancement, and color expansion into a single processing framework that can handle various image quality improvement tasks simultaneously, reducing redundant hardware requirements.
2Manufacturing precision
If multiple image processing algorithms are applied sequentially, then image quality parameters are improved, but processing time increases
Solution Approach 1:
The patent applies noise cancellation as the first processing step before any other image enhancement operations. By removing noise early in the processing pipeline, subsequent operations such as super-resolution and sharpness enhancement work on cleaner input data, which improves their efficiency and reduces the computational burden and time required for each subsequent processing stage.
Solution Approach 2:
The patent divides the image processing into distinct sequential stages, each handled by a specialized neural network. This segmentation allows each algorithm to focus on specific image quality parameters (noise, resolution, sharpness, contrast, color) independently, optimizing processing efficiency for each task while maintaining overall image quality improvement.
3Manufacturing precision
If neural networks are used for image processing algorithms, then image quality is improved, but hardware complexity increases significantly
Solution Approach 1:
The patent divides the complex image processing task into five specialized neural network modules, each responsible for a specific image quality parameter. This segmentation allows the system to use smaller, more efficient neural networks for each specific task rather than one large complex network, reducing overall hardware requirements while maintaining comprehensive image enhancement capability.
Solution Approach 2:
The patent implements a dynamic processing pipeline where the output of each neural network module serves as input to the next module in sequence. This dynamic architecture allows the system to adapt the processing flow based on the intermediate results, enabling efficient resource utilization and reducing unnecessary computational overhead in subsequent processing stages.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Disclosed is an image processing device according to an embodiment, including: a memory configured to store at least one instruction; and at least one processor configured to operate according to the at least one instruction to obtain a first differential image having a resolution greater than a resolution of an input image and a parameter value for a tone curve by applying the input image to a neural network, obtain a plurality of filtered images for different frequency bands by filtering the first differential image, apply, to the plurality of filtered images, gain values corresponding to a sample value of an edge map of a first image, obtain a second differential image by combining the plurality of filtered images, obtain a second image by combining the second differential image with the first image, apply, to the parameter value, a gain value corresponding to an average sample value of the edge map, determine a tone curve from the parameter value to which the gain value is applied, and obtain an output image by changing sample values of the second image according to the tone curve.